AI Agents On-Device: TraceWeave's $2.27M Phase-TITAN Bet

The most capable AI agents live in the cloud, but the most interesting places for them to work are everywhere the cloud isn't: the car, the drone, the factory floor, the glasses on your face. TraceWeave's thesis is that capable AI Agents On-Device are held back by the architecture, not the parameter count.

The seed round was announced on October 20, 2026, led by Mirae Asset Capital with participation from NAVER D2SF, according to the announcement carried by PR Newswire. The deal, worth about 3 billion won, valued TraceWeave at a pre-money 30 billion won (roughly $25.2 million), reported Dealroom. The company was founded in October 2025 by co-founder and CEO Jaeyoung Park alongside researchers with backgrounds at ETH Zurich, Google, and KAIST.

The memory bottleneck nobody solved by distilling

Most on-device AI work has focused on shrinking existing models — quantization, distillation, smaller parameter counts. TraceWeave is taking a more radical approach: replacing the Transformer architecture altogether with a proprietary design it calls Phase-TITAN. The problem it attacks is the key-value cache. In Transformer-based models, the KV cache grows with context length, so an agent holding a long reasoning chain needs steadily more memory. On a phone or a pair of smart glasses, that means strained memory, drained batteries, and heat. That is precisely the problem AI Agents On-Device face: the context that makes an agent useful is the same context that breaks the memory budget.

According to TraceWeave, Phase-TITAN keeps memory usage constant as context grows, because the company develops its own AI models rather than optimizing existing open-source ones. The contrast it cites is stark: at a 1-million-token context without quantization, a comparable Transformer's KV cache holds tens of gigabytes or more, while TraceWeave's 1-billion-parameter Phase-TITAN model holds just 0.04 gigabytes. As the announcement put it, removing the KV cache was never the hard part — staying capable without it was, and the company claims its architecture now delivers Transformer-class capability without what Park called the Transformer's "memory tax."

The numbers come with the standard caveats — they are company-run benchmarks. TraceWeave says its internal benchmarks on RULER and BFCL show Phase-TITAN maintaining performance on par with Transformer models of the same size and larger. Independent results from the broader BFCL suite are tracked publicly, per the company's own site, but agent builders will want third-party replication before treating the constant-memory claim as settled science.

From benchmark charts to the driver's seat

What separates this from pure architecture research is the traction. TraceWeave says it has demonstrated a multi-step agent running on a smartphone using only its CPU, with no cloud calls — agents that call tools, interpret results, and decide on next actions entirely on the device, a working proof point for AI Agents On-Device. In mobility, the company completed a paid pilot with the autonomous driving affiliate of a global automotive group, identifying driving-data frames that improved the partner's autonomous driving model without human annotation. It is now validating in-vehicle agent performance with a global automotive electronics software company, according to the NAVER D2SF announcement.

The same design is aimed at robots, drones, smart glasses, and wearables — environments where latency, cloud-call costs, and data privacy all push against round-tripping to a datacenter. TraceWeave was also a Top 20 finalist in NVIDIA's 2026 Inception Grand Challenge, reported by several startup trackers following the announcement.

Why agent builders should care

The strategic investor here matters. NAVER D2SF's head, Sanghwan Yang, said in the announcement that on-device AI is a core technology for running complex AI across device environments beyond PCs, and that it carries advantages in reducing cloud-call costs and protecting privacy — reported by the PR Newswire release. That framing points at where the agent economy is heading: agentic workflows where AI Agents On-Device need to act on sensor data in real time, work through network dead zones, or keep sensitive context local — the same enterprise push that has AI teams embedding deeper inside Global 2000 companies.

TraceWeave plans to use the seed funding to expand its research and engineering team and grow its business, operating out of NAVER D2SF's Gangnam base. The company combines engineering expertise grounded in computational neuroscience and physics with business development aimed at validating the architecture in real device environments, per the announcement.

For the agent community, the takeaway is architectural rather than financial: the memory tax of the Transformer is becoming the binding constraint on where agents can live. If post-Transformer designs like Phase-TITAN hold up under independent testing, the next wave of capable agents may not run in a datacenter at all — they may run as AI Agents On-Device with fixed memory, with the cloud as an optional accessory rather than a lifeline.